The industry has sorted its AI models into red and blue jerseys. ChatGPT is the Republican tool: 74 percent of GOP practitioners call it their most-used model. Claude is the Democratic tool: 82 percent of Democrats pick it first. Grok runs 45 percent on the right and 4 percent on the left. The DNC went further and barred its staff from ChatGPT and Claude entirely — Gemini is the approved model, in part because it integrates with the committee’s existing stack.

Read that as an operations question and it should bother you. Campaigns are choosing infrastructure the way voters choose cable news. The model your team uses to parse a county reject file is being selected by tribal signal.

The voter file doesn’t have a party. Neither does the work.

Where the model actually sits

Strip the punditry away and look at where a language model earns its keep in a turnout program. It is not on a debate stage. It fills seats — back-office seats, the kind nobody writes about.

The extraction seat: county files arrive as PDFs, scans, and forty different spreadsheet dialects. Something has to turn them into structured records before anyone can act on them. The matching seat: those records have to resolve against your file — this returned ballot is that voter, at this address, with this phone. The analysis seat: somebody has to write and maintain the code that ranks the work queue as new data lands. The drafting seat: per-county instructions, per-status variants, call scripts — drafted by machine, closed by a human, never shipped raw. And the QA seat: checking the other four.

Here is the thing about those seats. None of them ever asks the model its opinion. A model can hold whatever worldview its training gave it and still be the best extraction engine on the market, or the worst. The failure modes are boring and specific: hallucinated rows, missed matches, brittle code, a drafted instruction that cites the wrong county’s deadline. Different models fail differently at different tasks, and the rankings shift with every release cycle.

Which means choosing one brand for every seat — and choosing it by jersey color — guarantees exactly one outcome: you never benchmark half the market.

The map doesn’t even hold

If the tribal sorting reflected measured behavior, you could at least call it a crude heuristic. It doesn’t.

On June 24, The Washington Post published a test of the major models against a battery of political questions designed by researchers at Dartmouth and Stanford. ChatGPT — the Republican industry’s most-used tool — answered 80 percent of queries with only left-leaning arguments and offered a right-leaning-only answer 3 percent of the time. Grok, marketed to the right, cited left-leaning arguments more often than right-leaning ones. And the most balanced model in the test — offering both sides in over 90 percent of its answers — was Gemini. The one the DNC allows.

So the party that banned two models kept the most even-handed one, and the party that adopted ChatGPT as its workhorse is running the model that tested most one-sided against it. The sorting doesn’t track the testing. Which tells you what the sorting is: a signal, not a decision. Nobody ran the test. Everyone read the room.

There’s a second layer to the same survey data that gets less attention. Republicans aren’t just picking different models — they’re using them harder in some lanes and softer in others. GOP practitioners report using AI for image, video, and audio generation at more than double the Democratic rate, 61 percent to 25 percent. Content is where our side has planted the flag. The seats above — extraction, matching, analysis — belong to whoever bothers to sit in them.

What the build logs say

I came at this as a non-engineer. When I built Campaign Compass, the SaaS I shipped without a programming background, the build logs filled up with a pattern I didn’t plan: the model roster kept changing by task. The model that wrote the cleanest code was not the model I trusted to read a long, ugly file without dropping records. The model that drafted the best plain-English instruction was not the cheapest one to run a thousand times overnight. I stopped asking “which model do we use” and started asking “which model holds this seat” — and the answer moved every few months as new releases landed.

The method that fell out of that is a bake-off, and it costs an afternoon. Take one real task from your own program — say, last cycle’s reject file from your hardest county. Pull the same hundred records. Run three models against the same prompt. Count the errors by hand. Note the cost and the time. Now you have a number instead of a jersey. Repeat per seat, not per cycle — per release.

That’s not a research program. It’s the same discipline we’d apply to any vendor: diagnostic first, no tactic before the premise is proven. A model is a vendor. Treat it like one.

The margin math

Now the budget question, because this is ultimately a buying decision.

Run the arithmetic on a hypothetical 40,000-voter chase universe — mid-sized, nothing exotic. If two models differ by two points of accuracy in the matching seat, that is 800 voters mis-resolved: ballots credited to the wrong record, phones attached to the wrong household, doors knocked for people who already voted or letters mailed to people who never requested a ballot. Every one of those errors is money spent against a fiction, and most of them are invisible — the program “ran,” the reports filled in, the waste never surfaced. Turnout programs live or die on numbers smaller than 800.

Two points is not a claim about any particular model. It’s an illustration of what’s at stake in a decision most campaigns are currently making on vibes. The spread between the best and worst model on your specific files, in your specific seats, is an empirical question — and it’s answerable for the cost of one staffer’s afternoon.

So the buying question for any AI line item this cycle is not “whose model is it.” It’s this: which seat is this model filling, and what did it score in that seat, on our data? If the vendor can’t answer, the tool isn’t a capability. It’s merch.

The side that treats models as vendors — benchmarked, seat by seat, re-tested every release — will quietly out-execute the side that treats them as team flags. That edge won’t show up in a headline. It will show up where the margins live: in the reject file that got parsed right, the match that got made, the ballot that got chased on time.

1. AI is now on both sides of the signature line. Election offices already use non-generative AI to match mail-ballot signatures; on the campaign side, attorneys are running petition challenges through tools like Signafide, which indexes defective signatures line-by-line and cuts the review error rate from about 1.3% to 0.3% in roughly half the time (Parascript sells the verification side). For a chase program the lesson is the deadline, not the algorithm: the signature is where mail ballots quietly die, and the cure window — 24 states run a formal cure process, many with 24-hour rejection notice and a post-Election-Day clock — is the most under-worked deadline on the modern campaign calendar. Build the cure like you build the application drop: deadline-aware, named-voter, human-closed.

2. The left is consolidating “agentic” voter contact — and admitting the contact isn’t landing. Higher Ground Labs’ voter-contact showcase and its Agentic AI Open Call opened with a hard admission: in 2024 they reached more voters than ever and engagement still declined. The takeaway for the GOP: the edge isn’t a better bot at the top of the funnel — it’s the deadline-aware chase at the bottom that an agent can remind you about but can’t close.

3. “AI polls” are spreading — treat them as a hypothesis, not a return file. Aaru spins up ~5,000 synthetic voters and called the 2024 NY primary within about 371 votes, in minutes, at a tenth of survey cost — and skeptics warn you can simply spin up more bots to “fake-poll” a small race into telling you what you already believe (Nate Silver’s shop is blunter still). Useful to pressure-test a Universe hypothesis before you spend; never a substitute for the live return file your chase is actually scored against.

4. Ad creative is being commoditized; the margin moves to distribution. OpenAI is rolling out tools to generate, modify, localize and translate ad creative, and the IAB reports 86% of media buyers use or plan AI-generated video in 2026. When everyone can make the spot in an afternoon, the spot stops being the advantage. The margin moves to which universe sees it — the same diagnostic discipline that decides which ballots you chase.

The through-line: AI keeps getting better at the parts of the job that scale (drafting, simulating, matching) and no better at the part that wins — closing a named voter against a real deadline.

That’s still human work.

Field is the last leg of the program still priced in human hours. That is exactly where the AI money belongs — and it is not on the doorstep.

I want to be precise about that, because the loudest pitch in field right now is the opposite one. There’s a whole new category of “agentic” door-knocking products promising to replace the canvasser with a voice agent — one vendor is literally marketing a system it says “ends candidate canvassing in 2026.” That’s the wrong end of the problem. The knock is the single most effective tool in turnout. You don’t automate away your best instrument. You automate the day of labor stacked around it.

The mechanics: what a knock actually costs

Start with why the knock matters. The research here is old and settled. Across the field-experiment literature, in-person canvassing is the strongest get-out-the-vote tactic there is — stronger than phone, far stronger than mail. Green, McGrath and Aronow’s precision-weighted estimate puts the turnout effect of face-to-face canvassing around 2.5 points. The canonical rule of thumb from Green and Gerber’s Get Out the Vote is roughly one additional voter for every fourteen doors where a canvasser actually has a conversation. Mail, by contrast, moves turnout by a fraction of a point per piece.

So the knock is scarce and precious. A canvasser gets through a finite number of doors in a shift, converses behind fewer, converts a fraction of those. That conversation is the asset. Everything else in field exists to point it at the right door.

And “everything else” is a pile of labor nobody writes about. Building the door universe. Cutting it into balanced turf. Sequencing each walk so canvassers aren’t crisscrossing a precinct. Printing and syncing lists. Writing the per-door instruction. And then — the part that actually bleeds margin — re-cutting all of it every time the return file moves. Because in a mail-heavy electorate, the field universe is a moving target. Ballots come back daily. Every door on an active turf whose ballot is already in the box is now a dead door: a knock spent on a vote you already have.

Here’s the margin math nobody runs. Say a canvasser knocks eighty doors a shift and your walk list is three days stale. In a competitive vote-by-mail environment, a meaningful slice of those doors returned their ballots since the list was cut. Those knocks didn’t move turnout — they confirmed it. You paid your most expensive, most effective turnout dollar to talk to people who were already done. Multiply by every canvasser, every shift, across the whole program, and the leak is enormous and invisible, because it still reads as “we knocked our doors.”

This is where AI changes the economics — on the labor, not the doorstep. Building a door universe from a plain-language description instead of a data tech’s afternoon. Cutting and balancing turf in minutes. Sequencing routes against density and distance — the AI-driven routing now shipping in field apps like Vizit and Ecanvasser auto-adjusts walk paths from real canvasser progress. Drafting the canvasser’s per-door note. And, most important, suppressing returned-ballot doors against the live return file every single night, so the list in a volunteer’s hand the next morning points only at outstanding ballots. None of that replaces the conversation. All of it makes sure the conversation lands somewhere that matters.

A worked example, at scale

Run the numbers on a real ground operation. In 2024 we turned out 650,000 voters for President Trump in Georgia as the state’s America PAC director. At that scale the binding constraint is never strategy. It’s the labor of keeping thousands of walk lists true against a return file that moves every day across a whole state.

Picture the old way. A county’s mail returns post. Somewhere in that universe are thousands of doors whose ballots just arrived. The data team re-pulls, re-cuts, re-prints — on a lag of days, because that’s how long the human pipeline takes. For those days, canvassers walk dead doors. Not because anyone was sloppy. Because the labor of re-cutting a statewide field universe by hand simply cannot keep pace with a daily return file.

Now picture the AI way on the same operation. The return file lands at night. A script reconciles it against every active turf, suppresses the doors whose ballots are in, re-balances the walk lists, and regenerates routes before the morning launch. Canvassers walk a list that is true that day. The same number of knocks, the same volunteers, the same budget — pointed only at the ballots still outstanding. That is not a better doorstep. It’s the same doorstep, aimed correctly, every day instead of every third day.

And you don’t need a vendor’s roadmap to build it. I shipped Campaign Compass — chase tooling — as a non-engineer working with AI. The walk-list re-cut is, under the hood, a reconciliation-and-routing problem: match the return file to the turf, drop the closed doors, re-sequence, export. That is now buildable by an operator who actually understands the field workflow, not only by a software company that has never cut turf. The people who know where the margin leaks are now close enough to the tools to plug it themselves.

So what — for a campaign budgeting right now

If you’re a GOP campaign making field decisions this cycle, here’s the takeaway. Your canvass is your most expensive turnout dollar and your most effective one. The temptation is to spend the AI budget making the doorstep cheaper — a voice agent instead of a volunteer. Resist it. An agent can leave a reminder; it cannot be the neighbor on the porch, and the porch is the part the research says works. Spend the AI budget collapsing the hours between the return file moving and the walk list changing.

There’s a single buying question that exposes whether a field operation is leaking margin: When a batch of ballots comes back, how long until my canvassers stop knocking those doors? If the honest answer is “next print run” or “next week,” you are paying your best people to talk to voters who are already finished. The cheap win in field isn’t a smarter door. It’s a list that’s true this morning.

That’s the whole discipline, really. Campaigns are won on the margins, and field is where the largest, least-examined margin still sits — not in the conversation at the door, but in the day of stale labor standing between the data and the door.

This week the voter file started answering in plain English, the persuasion research got uncomfortable, and the courts drew a line. Four notes from inside the work.

1. The voter file now answers in plain English. IQM launched Custom Voter Audiences on June 11 — a natural-language layer over a verified voter file that lets you describe a universe (”low-propensity, GOP-leaning mail voters in these three counties”) and get an activatable segment in seconds instead of un-checking boxes in a data table (PRNewswire; AdExchanger). For us this is universe construction at the speed of conversation — but the segment is only as honest as the file and the propensity model underneath it, and “turnout propensity” described in a sentence is still a model output you validate against the live return file, not a fact.

2. The chatbot persuasion numbers are real, and they cut both ways. Two peer-reviewed studies (Nature, Science) found conversational AI moved voters more than conventional ads — one pro-Harris model shifted likely Trump voters 3.9 points, roughly four times the effect of tested 2016–2020 ads (Cornell Chronicle; MIT Technology Review; Scientific American). The operator lesson isn’t to build a persuasion bot; it’s that the low-information voters in your chase universe are getting election facts from these tools, so verify what the chatbot tells them about deadlines and drop-off before it costs you a ballot.

3. Agentic voter contact is consolidating into a category. Between Colloquy’s automated voice/SMS GOTV agents and the new wave of “agentic” conversational door-knocking systems being pitched for 2026, the automated-contact stack is no longer a novelty (Campaigns & Elections; The AI Journal). Worth a hard look for top-of-funnel reminders — but ballot chase is deadline-aware human follow-up on cure and signature defects, and that’s exactly the work an agent can’t close.

4. The courts settled on “disclose, don’t ban.” Maryland became the 30th state with an election-deepfake law (effective June 1), while the ban at the heart of California’s AB 2839 stayed permanently enjoined after Kohls v. Bonta — leaving disclosure-style rules standing and outright bans on shaky ground (Public Citizen; Biometric Update; Broadcast Law Blog). If you’re deploying any synthetic media this cycle, build to the disclosure deadline the way you build to the mail-ballot deadline: it’s a date on the calendar, and missing it is unforced.

The margins are where these tools either earn their keep or quietly cost you ballots. We read this stuff so your program doesn’t get surprised.

Subscribe free to Winning on the Margins → margins.catoconsultinggroup.com

There is a universe of voters who did everything you asked. They requested the ballot. They filled it out. They signed it and sent it back. And right now a county clerk is getting ready to throw some of those ballots in the reject pile over a signature that doesn’t match a twenty-year-old DMV card — unless someone reaches the voter in the next few days and walks them through the fix.

That list has a name. It’s the cure universe. It is the smallest turnout universe in the whole program, and ballot-for-ballot it is the cheapest margin on the board. Almost no Republican campaign has a single person assigned to it.

That’s the opportunity. Let me show you the mechanics, then where AI changes the math.

THE MECHANICS NOBODY WRITES ABOUT

Start with the numbers, because they set the size of the prize. The nationwide mail-ballot rejection rate in 2024 was 1.2 percent — higher than 2020 and higher than 2016. The single most common reason a ballot got rejected was a non-matching signature: 40.7 percent of all rejections nationally, and in some states the leading reason by a far wider margin. In California, an analysis found 59 percent of rejections came down to signature.

Now the part that matters strategically. Of the ballots that hit a problem and entered a cure process in 2024, only 54.3 percent were actually fixed in time. Nearly half of a confirmed, motivated, already-returned set of voters simply fell off the table — not because they changed their minds, but because nobody got to them before the clock ran out.

And the clock is the whole game. Cure is the most deadline-aware mechanic in mail. The rules are different in every state and sometimes every county: some let a voter cure up to a date certain after the election, some count days from the moment the ballot was flagged, some require a signed affidavit, some a copy of ID, a few want a notary. The voters who get caught are disproportionately the youngest and the newest — 18-to-24-year-olds had more than 3 percent of their mail ballots rejected in 2024, most often on signature. In one Washington county, voters 18 to 26 were 51 percent of all signature challenges. These are not your hardest persuasion targets. They are people who already voted.

So the cure universe has a strange profile. Every member is a confirmed voter. The universe is tiny — call it 1.5 percent of returns. It decays faster than anything else you touch, because each voter has a personal deadline measured in days. And the upside is enormous, because the research is now unambiguous that working it moves the count. A University of Pennsylvania study published this month, covered by Votebeat, found that voters who were actually notified about a problem with their ballot were 25 percentage points more likely to end up with a counted vote. In Pennsylvania, counties that allow curing ran a 0.49 percent rejection rate against 0.59 percent in counties that don’t — roughly 17 percent fewer voters thrown out, from the policy alone.

So why does almost nobody run a real cure program? Because operationally it’s miserable.

WHY THE WORK DOESN’T GET DONE

A cure operation is a nightly race against a moving target. Counties post rejections in rolling batches, in whatever format they please — a portal here, a daily PDF there, a CSV with column headers that change between counties. Someone has to pull each of those, match the rejected names back to your voter file, find a phone number that still works, figure out how many hours that specific voter has left under that specific county’s rule, decide whether this is a call, a text, a door, or a notary problem, and then write the voter the exact instruction for the fix — which is different in the next county over.

Do that for a 343,000-voter chase program and the cure tail alone is thousands of ballots, arriving in dribs and drabs, every one on its own clock, in the last week when your team is already drowning. That is exactly the work that gets a polite “we’ll get to it” and never gets gotten to. The economics never penciled out: the universe was too small to justify standing up a war room, and too manual to run on the side.

That last sentence is the one AI rewrites.

WHERE AI CHANGES THE ECONOMICS

Take the four miserable steps and look at each as a place the cost just collapsed.

Ingestion. The rolling county files — the portals, the inconsistent PDFs, the CSVs that never match — are now a parsing problem a model handles in seconds instead of a volunteer handles at 9 p.m. You point it at the day’s batch and it returns clean, normalized rejected-voter records on the program’s cadence, not whenever someone has time.

Matching. Tying a rejected name back to your file, with a reachable phone, is entity resolution — the same capability that’s already reordering voter-file work. It’s cheap now.

Triage. This is the real unlock. The cure list isn’t a list, it’s a queue ordered by hours-to-deadline crossed with modeled support crossed with how reachable the voter is. A model can re-rank that queue every time a new county batch lands, so your callers are always working the ballot that is both most savable and closest to dying — instead of working alphabetically and losing the ones at the bottom.

Instruction. The fix differs by state and county — affidavit, ID copy, notary, drop-off versus mail-back. Generating the correct, plain-language cure script for each voter’s exact situation is a drafting task, and drafting is the thing these tools do best. The human stays on the phone, where humans belong. The lookup, the paperwork logic, the per-county rulebook — that moves to the machine.

None of this is persuasion. Nobody’s mind is being changed. This is taking a confirmed vote that the system is about to discard and not discarding it.

THE “SO WHAT” FOR A CAMPAIGN BUDGETING RIGHT NOW

Here’s the diagnostic question for anyone building a 2026 mail budget: when your program produces rejected ballots — and it will — does anything in your stack catch them, rank them against their deadlines, and put a reachable script in a caller’s hand the same day? For almost every Republican campaign the honest answer is no. The cure tail is treated as breakage, a rounding error in the GOTV report.

It isn’t a rounding error. It’s the highest-conversion universe you will ever touch, because every voter in it already tried to vote for you. Half of them are currently being lost for want of a phone call. We defeated a $450 million school bond by 319 votes against seven-to-one money. In a race that size, the uncured pile is not a footnote. It is the margin.

The firms that win the close ones in 2026 won’t be the ones with the cleverest synthetic-voter model. They’ll be the ones who used AI to make a job that never penciled out — working every flagged ballot to its deadline — finally pencil out. Diagnostic first, deadline-aware, executed in-house. The margin was sitting in the reject pile the whole time.

The case in one paragraph. Argued March 23, 2026, Watson turns on whether the federal statutes setting “the first Tuesday after the first Monday in November” as Election Day require ballots to be received by that date, or merely cast. Mississippi counts ballots postmarked by Election Day that arrive within five business days. The RNC and its co-plaintiffs argue federal law preempts that window. At oral argument, a majority of justices appeared sympathetic. Counsel for the challengers told Justice Kavanaugh a June ruling would not arrive too late to implement, since November mail ballots will not yet have been sent.

The stakes, in numbers. At least 750,000 grace-period ballots arrived after Election Day in 2024 and were counted, by Votebeat’s analysis. Lateness is already the single largest reason mail ballots get rejected — roughly 103,000 in 2024, about 18% of all mail-ballot rejections, per the EAC’s Election Administration and Voting Survey. And this is not a blue-state question: Alaska, Mississippi, Texas, and West Virginia all keep grace periods on the books. If the Court rules for the plaintiffs, the change lands on Republican programs too.

What the data actually says. Here is the number that should drive your planning. States with and without grace periods reject mail ballots for lateness at nearly identical rates — roughly 0.2 percent of returned ballots, a statistically insignificant difference, per the Bipartisan Policy Center’s analysis of EAVS data. Voters adapt to whatever deadline they are given. The operational lesson follows directly: programs should adapt before the deadline moves, not after. The grace period is a buffer, not a plan.

The diagnostic view. In our framework, this is the Landscape line of inquiry — the statutory environment the program must live in. If the diagnostic isn’t honest about the calendar, the strategy isn’t useful. Right now, an honest diagnostic in any grace-period state says: the receipt deadline for federal races may move to Election Day mid-cycle, with a ruling that could be effective immediately or prospectively. You do not get to know which. So you build the program that wins in both worlds.

What a deadline-aware program does now. Build the chase calendar to an Election Day receipt deadline regardless of the ruling. Pull the final mail-return cutoff up a week. Schedule the final-week flip from mail return to in-person delivery earlier, and script it into every chase channel. Write one instruction into every script — returned and received by Election Day — because if grace periods fall for federal races while state-race windows linger on the books, voters in those 14 states face split deadlines, and a chase program that needs a footnote to explain itself is leaking margin. Treat military and overseas voters as their own track: UOCAVA ballots are rejected for lateness at more than eight times the domestic rate. And do not lean on postmarks. USPS itself says postmarking was never designed as proof of mailing date.

The margin is in the discipline. Late-arriving ballots are a fraction of the vote — between 0.1 and 2.5 percent of mail ballots in reporting states. But margins are where modern races are decided, and a program that banks its ballots three weeks early is indifferent to whatever the Court does this month. We build programs around the statutory environment as it is, not as it was. This summer, “as it is” may change in the middle of the cycle. That is not a reason to wait for the ruling. It is the reason to build now.

Two campaigns in the same district buy the exact same voter file from the exact same vendor on the same Tuesday. By Election Day one of them has wasted fifteen percent of its mail budget and the other hasn’t. The file didn’t change. What changed is that one of them treated the file as a finished product and the other treated it as raw material.

This is the part of a turnout program nobody writes about. The conversation about AI in campaigns is all happening at the glamorous end — synthetic electorates, message testing, the poll that never sleeps. Meanwhile the single highest-leverage thing AI does for a turnout program is the least glamorous task in all of politics: cleaning the list. Not modeling who’s persuadable. Just making sure the name, the address, and the status on each record are true before you spend a dollar acting on them.

Here’s why that matters more than it sounds. A mail-and-chase program is a machine for spending a fixed budget against a list of human beings, on a deadline that does not move. Every record in that list is an instruction: send this person a ballot application, mail this person a reminder, route a knock to this door, point a phone call at this number. If the record is wrong, the instruction is wrong, and the money is gone — not misallocated, gone. A mail piece to someone who moved two years ago doesn’t underperform. It does nothing. A chase call to a landline that was disconnected in 2023 isn’t low-yield. It’s zero-yield. And here’s the trap: none of it shows up as failure. It shows up as “we ran the program.” The waste is invisible because the dead records sit quietly in the universe looking exactly like the live ones.

Now the scale of the rot, with real numbers. Americans move constantly — about 25.9 million people changed residences in 2024, roughly 11.8 percent of the population, and that’s at a record low mover rate. One in nine of the people on your file moved since the last cycle, and a meaningful share moved across a precinct, county, or state line that breaks their registration. Layer death on top: this April, North Carolina alone identified roughly 34,000 deceased individuals still sitting on its voter rolls after a single database comparison. The list-maintenance machinery that’s supposed to catch this — NCOA runs, Social Security death matches, the ERIC cross-state checks used by some two dozen states — is real but it is monthly, lagged, and uneven from jurisdiction to jurisdiction. The file you buy is always stale at the edges. The only question is whether you find the stale records before you mail them or after.

This is where AI changes the economics, and it has nothing to do with chatbots. The hard part of file hygiene was never knowing the rules — everyone knows a deceased flag matters. The hard part is resolution at scale: deciding whether “Robert J. Smith, 412 Oak” and “Bob Smith, 412 Oak St Apt 2” are one voter or two, whether a forwarding address is a real move or a snowbird’s winter rental, whether two records that share a household are duplicates or a father and son. That’s entity resolution, and it used to require either an expensive enterprise data vendor or a junior staffer burning a week on a spreadsheet doing it by eye. Language models are unusually good at exactly this fuzzy-matching, normalize-the-mess, judge-the-edge-case work. The voter-file industry knows it — the 2026 file refresh added a forward-looking “Moved” flag specifically to catch voters matched to a new registration in another state, and match-rate quality (how cleanly your file maps to real, reachable people) is now the thing serious data shops compete on. The leverage isn’t that AI finds voters you didn’t have. It’s that AI lets a small team scrub a universe down to the records that are actually true, in hours, for the price of compute instead of the price of a data contract.

I’ve built both sides of this. Running Georgia as a state director for America PAC, we were responsible for a turnout operation that ultimately delivered 650,000 voters for President Trump. At that scale the file underneath the program isn’t a detail — it’s the program. Every percentage point of dead, moved, or duplicate records is mail printed for nobody and field hours pointed at empty houses, multiplied across hundreds of thousands of targets. You feel the cost of a dirty file in your gut long before it shows up in a report, and you learn to fix the file first, because everything downstream inherits its errors.

That conviction is why I built Campaign Compass — a piece of software I shipped myself, as a non-engineer, by writing the build logs in plain English and letting AI handle the code. A lot of what Compass does is exactly this unglamorous resolution work: take the raw file, find the duplicates, flag the movers, normalize the addresses, and hand the operator a universe they can actually trust before they spend against it. The point of telling you a non-engineer built it is the point of this whole essay: the file-hygiene work that used to be gated behind a vendor’s enterprise contract is now something a disciplined operator can do directly, because the AI does the part that used to require a specialist. The moat around clean data is draining, and it’s draining toward whoever bothers to walk through it.

So here’s the “so what” for anyone setting a 2026 turnout budget. Before you spend a dollar on a clever AI persuasion tool, ask what you’re spending on the file the whole program runs on. The order of operations matters: a brilliant chase strategy executed against a fifteen-percent-rotten universe is worse than a plain strategy executed against a clean one, because the dirty file taxes every single touch you make. Run the math on your own program. If even ten percent of your file is dead, moved, or duplicated — a conservative read given that one in nine voters moves every year — then a tenth of your mail spend, your chase hours, and your field walk is being burned on records that can never produce a vote. In a race decided by a few hundred ballots, that is not housekeeping. That is the margin. We beat a $450 million bond by 319 votes by being disciplined about exactly this kind of thing. The buying question is simple: before this tool helps me reach voters, does anything in my stack make sure the voters on my list are real? If the answer is no, that’s the first check you should write, and it’s the cheapest one on the board.

Winning on the Margins is where I write about the parts of campaigns nobody else does — mail ballot, ballot chase, the voter file, and field. If you build or fund turnout programs, subscribe here: https://margins.catoconsultinggroup.com

They cover the industry. They speculate about what the tools might do. They write the trend piece, run the panel, post the hot take. Very little of it is written by anyone who has actually built a turnout program, signed the vendor contracts, owned the data, and watched the numbers come in on election night.

That’s the seat I’m writing from. And it’s the gap this newsletter exists to fill.

Subscribe now

Here’s the premise, stated plainly: campaigns are won on the margins. Not on the loudest ad or the best-funded air war — on turnout mechanics. Mail ballot. Ballot chase. The voter file. Field. The unglamorous, deadline-aware execution that decides close races. It’s the discipline Democrats spent two decades perfecting and most conservative campaigns still underinvest in. And it is exactly where AI is about to change the economics of everything.

Not in the “AI will transform democracy” sense. I have no interest in that conversation. I mean something narrower and more useful: which voters you can identify, how fast you can build a chase universe, what it costs to cure a ballot, how many doors one organizer can effectively run, how quickly you can turn a raw file into a targeting plan. Those are mechanical questions with real dollar figures attached, and AI moves those figures right now. That’s what I want to write about.

A little about the seat, so you know the receipts are real. I built and ran a ballot-chase program across a 343,000-voter universe in Sedgwick County, Kansas. I served as Georgia State Director for America PAC, where 650,000 voters turned out. I helped defeat a $450 million school bond by 319 votes while being outspent seven to one by the largest school district in the state of Kansas. And I shipped Campaign Compass, a campaign-operations SaaS — as a non-engineer, built with exactly the tools I write about here. Fifteen years of underdog wins, most of them decided at the margin.

Everything in this newsletter will be grounded in work that was actually done, with numbers you can check — not predictions, not vibes. When I make a claim, I’ll show you where it comes from. When something is my analysis rather than established fact, I’ll say so. When a number can’t be verified, I say so upfront.

This isn’t a partisan rant, either. It’s written from a Republican seat, for people who want the Right to stop leaving votes on the table — but the mechanics don’t care about your politics, and neither will the analysis. The other side built a turnout machine while a lot of our side was still arguing about whether mail ballots were legitimate. I’d rather we just win the argument by winning the race.

Here’s what you’ll get, and how often:

Every Tuesday — one original argument about applying AI to turnout mechanics. A single idea, worked all the way through: what actually happens in the workflow, where AI changes the math, a concrete example, and what it means for a campaign making budget decisions right now.

Every Friday — field notes. Three to five short items: what shipped this week, what changed in the rules or the tooling, what the other side is building, each with one line of operator commentary on what it means for GOP turnout work. Lighter, faster, and the fastest way to stay current without drowning in the firehose.

That’s it. Two emails a week, both built to be read on your phone between meetings, both written for people who run campaigns for a living rather than people who tweet about them.

If that’s you — if you’re a campaign manager, a PAC director, a candidate, or an operative trying to figure out what’s real and what’s hype before you spend money on it — this is for you.

The margin is where elections are decided. It’s about to be where the AI advantage is decided too. I’d like to take you along for it.

Subscribe free, and you’ll get the first Tuesday essay in your inbox this week.

— Ben

Benjamin Davis, founder, Cato Consulting Group